Page 54 - Read Online

Page 457 Goyal et al. Art Int Surg. 2025;5:450-64 https://dx.doi.org/10.20517/ais.2024.108

Table 3. Comparison of AI models, annotation tools, and accuracy metrics for surgical phase recognition

Number of F1
Authors Annotation software AI model type Model architecture Overall accuracy Precision
surgeons score
Ortenzi et al. Manual annotation Surgical Intelligence Platform Video Transformer Network (VTN), Long Short- 85 88.8% (89.16% with an overall 89.59% for - -
[10]
(2023) (Theator Software Co.) Term Memory (LSTM) network (for model unilateral; 83.17% with an overall 88.45%
adaptation) for bilateral)
Takeuchi et al. CVAT (Computer Vision DETR (Detection Transformer) - 11 77.10% 75.40% 51.20%
[20]
(2023) Annotation Tool)
Choksi et al. Manual annotation ResNet-50 (Microsoft - 8 68.7% (59.8% - 78.2%) - -
[21]
(2023) Research)
Takeuchi et al. Indexiv Annotation TeCNO (Transsion Holdings) & Temporal Convolutional Network (MS-TCN) + 11 Unilateral: 88.81%; Bilateral: 85.82% - -
[8]
(2022) Platform Hidden Markov Model (HMM) Hidden Markov Model (HMM)
Zang et al. Manual annotation CV/DL ResNet-50, Video Swin Transformer, Perceiver 8 85% 75.81% -
[9]
(2023) IO
Zygomalas et al. SuperAnnotate v1.1.0 YOLOv8 (Ultralytics Inc.) - 5 87.30% 82% > 50%
[7]
(2024)

AI: Artificial intelligence.

These models were trained using annotated datasets, allowing the systems to learn and identify specific structures accurately. The choice of model architecture

was influenced by the nature of the input data (videos or images) and the specific requirements of the task, such as temporal modeling, feature extraction, or
real-time analysis.

Surgical phase recognition

Significant heterogeneity exists in the identification and classification of surgical phases during IHR across included studies. While some studies focus on
broad procedural steps such as “balloon dissection” and “mesh placement” , others incorporate a more granular breakdown, identifying anatomical
[10]
landmarks like the “pubic symphysis” and “iliopsoas muscle” . Additionally, phases related to peritoneal manipulation, including “peritoneal flap incision”
[20]
and “closure,” were emphasized in works by Takeuchi et al. and Zang et al. .
[9]
[8]

Notably, certain steps critical to surgical workflows, such as “needleholder” handling and recognition of vascular or ligamentous structures, were unique to
specific models . Late procedural phases, including “stationary idle” and “blurry” recognition, were inconsistently reported, underscoring variability in
[7]
endpoint definition [Table 4]. Establishing a standardized framework for phase segmentation remains critical to improving model performance, ensuring
comparability, and advancing automation in inguinal hernia surgery.